An AI multi-agent and digital twin fusion production process visualization method, medium and system

Abstract

The application provides an AI multi-agent and digital twin fusion production scheduling process visualization method, medium and system, belonging to the technical field of AI multi-agent production scheduling. Large-scale parallel computing is realized by constructing a GPU three-layer CUDA processing architecture. The first layer of data preprocessing grid performs data cleaning in parallel. The second layer of negotiation analysis grid runs a Transformer-based agent interaction recognition model for parallel pattern recognition. The third layer of visualization calculation grid runs an LSTM-CNN fusion production scheduling process mapping model to generate visualization data. The multi-head attention mechanism dynamically adjusts the allocation of computing resources. The pipeline parallel and data parallel strategies optimize the computing performance. The asynchronous data transmission and computing overlap technology hides the memory access delay, solving the technical problem that multi-agent high-frequency negotiation data cannot be processed in real time.

Description

Technical Field

[0001] This invention belongs to the field of AI multi-agent scheduling technology. Specifically, it relates to a method, medium, and system for visualizing the scheduling process by integrating AI multi-agent and digital twin. Background Technology

[0002] In the field of intelligent manufacturing, traditional production scheduling systems primarily rely on CPU serial processing architectures to handle multi-agent negotiation data. They sequentially process data such as equipment agent bidding information, order agent demand parameters, and material agent inventory status using single-threaded or multi-threaded methods, employing traditional data processing pipelines for data cleaning, feature extraction, and pattern recognition. However, as multi-agent systems scale and negotiation frequencies increase, the performance bottleneck of traditional CPU processing architectures becomes increasingly apparent when dealing with 1000Hz high-frequency sampled multi-agent negotiation data. Data processing delays severely impact the real-time performance of scheduling decisions. Existing technologies generally employ time-sharing strategies, processing complex negotiation data in batches, resulting in low data throughput and an inability to meet the computational demands of real-time negotiation analysis, particularly when handling large-scale agent interaction recognition and scheduling process mapping calculations. In current multi-agent scheduling systems, the computational limitations of traditional processing architectures prevent real-time parallel processing of frequently generated negotiation data, leading to delays in negotiation behavior recognition and visualization data generation, severely impacting system response speed. In other words, existing technologies suffer from the technical problem of being unable to achieve real-time parallel processing of high-frequency multi-agent negotiation data. Summary of the Invention

[0003] In view of this, the present invention provides a method, medium and system for visualizing the scheduling process by integrating AI multi-agent and digital twin, which can solve the technical problem in the prior art that high-frequency negotiation data of multi-agent cannot be processed in real time in parallel.

[0004] The present invention is implemented as follows: The first aspect of the present invention provides a method for visualizing the production scheduling process by integrating AI multi-agent and digital twin technologies, comprising: establishing a distributed data acquisition layer; collecting workshop equipment status data, material location data, and multi-agent negotiation data through a 5G sensor network; using a Kalman filter to remove sensor noise from the workshop equipment status data, material location data, and multi-agent negotiation data; converting heterogeneous data into a unified spatiotemporal data format and storing it in a time-series database; activating a three-layer CUDA processing architecture of the GPU; the first layer, a data preprocessing grid, responsible for real-time data cleaning and feature extraction of the spatiotemporal data format; the second layer, a negotiation analysis grid, running an agent interaction recognition model to process the multi-agent negotiation sequence in the multi-agent negotiation data; and the third layer, a visualization computing grid, running a production scheduling process mapping model to generate visualization data; constructing a multi-agent negotiation state graph; and using a force-directed placement algorithm to arrange agent nodes according to their bid prices. Size is mapped to node radius, and similar interactive behaviors are grouped and displayed using a dynamic clustering algorithm. Combined with a timeline controller, the negotiation process is replayed in segments. A dual-engine rendering system is established, synchronizing virtual and real-world rendering. The real-time rendering engine uses a dynamic LOD algorithm to adjust display precision based on viewing distance, while the offline precision calculation engine generates replay data based on ray tracing technology. A spatiotemporal indexing mechanism enables data linkage between the two engines. An anomaly propagation visualization early warning mechanism is employed, constructing an order resource dependency graph. Dijkstra's algorithm is used to calculate the shortest propagation path of anomaly impacts, and a particle system is used to simulate the diffusion process. An adaptive view recommendation algorithm is implemented, collecting user operation data to construct feature vectors and using a personalized view recommendation model to output the optimal view combination scheme. Production scheduling decisions are optimized based on visualized data, integrating multi-agent negotiation transparency, anomaly propagation early warning information, and user view feedback into production scheduling adjustment parameters to generate the final production scheduling scheme and output it to the workshop execution system.

[0005] The 5G sensor network collects data at a frequency of 1000Hz and establishes a hierarchical processing system for multi-scale negotiation matrix mapping. The hierarchical mapping from the device-level scale to the workshop-level scale enables multi-level extraction of multi-agent negotiation sequences.

[0006] The agent interaction recognition model is a lightweight neural network based on the Transformer architecture, consisting of a 4-layer encoder and a 2-layer decoder. It employs a multi-head attention mechanism to process multi-agent negotiation sequences. The number of attention heads is dynamically adjusted based on the current number of active agents, calculated as: Number of attention heads = Number of agents / 8 + 2, with a maximum limit of 16 heads. The number of model parameters is controlled within a certain range. Within.

[0007] The steps for establishing the training dataset for the intelligent agent interaction recognition model include collecting 12 months of historical production scheduling logs, extracting intelligent agent ID, interaction timestamp, bidding parameters, and negotiation result fields to construct a multi-agent negotiation sequence, segmenting the data according to a 300-second time window, labeling abnormal and normal negotiation events, and including positive samples in the dataset. negative samples The negative samples were augmented using data augmentation techniques. strip.

[0008] The training steps of the intelligent agent interaction recognition model include using the AdamW optimizer, setting the learning rate to 0.001, the batch size to 32, the training rounds to 100, using the cross-entropy loss function, evaluating the model performance on the validation set every 10 rounds, and triggering an early stopping mechanism when the validation accuracy does not improve for 5 consecutive rounds.

[0009] The production scheduling process mapping model is a time-series data processing model based on convolutional neural networks. It integrates LSTM for processing time-series data and CNN for processing spatial relationship data. It uses an attention fusion module to combine multi-dimensional features. The spatial attention weight is determined based on three parameters: equipment importance coefficient, material flow frequency, and probability of anomaly occurrence.

[0010] The steps for establishing the training dataset for the production scheduling process mapping model include collecting historical production scheduling data from 50 workshops, including equipment operating parameters, material flow records, and process execution sequences. Training labels are then constructed using real-time digital twin status data and categorized into three levels—simple, medium, and complex—based on scheduling complexity, with each category containing data samples. indivual.

[0011] The production scheduling process mapping model training steps include a phased training strategy: the first phase pre-trains the LSTM temporal processing module for 50 rounds; the second phase jointly trains the CNN spatial processing module and the LSTM temporal processing module for 100 rounds; and the third phase fine-tunes the attention fusion module for 30 rounds, using the Adam optimizer with an initial learning rate of 0.01, which decays to 0.1 times the original rate every 30 rounds.

[0012] The multi-agent negotiation data includes five types of data: equipment agent bidding information, order agent demand parameters, material agent inventory status, scheduling agent coordination instructions, and quality agent detection results. Each type of data includes a unique agent identifier, timestamp, status code, parameter value, and priority weight field.

[0013] The multi-agent system includes an equipment agent responsible for equipment status monitoring and capacity bidding, an order agent responsible for order demand analysis and priority management, a material agent responsible for inventory monitoring and delivery scheduling, a scheduling agent responsible for resource allocation and conflict coordination, and a quality agent responsible for quality detection and anomaly warning.

[0014] The force-directed layout algorithm refers to a graph layout algorithm based on physical simulation, which treats the intelligent agent nodes as point masses and the lines between nodes as springs, and achieves a balance state by iteratively calculating the attraction and repulsion between nodes.

[0015] The dynamic clustering algorithm refers to the improved K-means clustering method, which dynamically determines the number of cluster centers based on the interaction frequency and behavioral similarity of agents, and groups agents with similar negotiation behaviors into one class.

[0016] The dynamic LOD algorithm refers to an algorithm that dynamically adjusts the level of detail displayed based on the observer's viewpoint distance. Full-precision data is displayed when the viewpoint distance is ∈ (0, 5]m, medium-precision data is displayed when the viewpoint distance is ∈ (5, 20]m, and simplified data is displayed when the viewpoint distance is > 20m.

[0017] The Dijkstra algorithm mentioned here refers to a classic algorithm for calculating the single-source shortest path in a graph. It is used to quickly find the shortest propagation path of anomaly impact in an order resource dependency graph, with a time complexity of O(n log n). .

[0018] The personalized view recommendation model refers to a recommendation system based on a combination of collaborative filtering and deep learning, which predicts the most suitable combination of visualization views by analyzing the user's historical operation behavior.

[0019] In the adaptive view recommendation algorithm, when the sampling frequency is ∈ [1, 120) Hz, the cubic spline interpolation reconstruction algorithm is started to avoid view switching distortion. When the sampling frequency is ≥240 Hz, low-pass filtering downsampling is used to reduce data redundancy.

[0020] A second aspect of the present invention provides a computer-readable storage medium storing program instructions, which, when executed in a computer, are used to perform the above-described method for visualizing the scheduling process by integrating AI multi-agent and digital twin technologies.

[0021] A third aspect of the present invention provides a production scheduling process visualization system integrating AI multi-agent and digital twin, comprising the aforementioned computer-readable storage medium. The system is any one of a computer, a server, or a microcontroller. The computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes the program instructions stored in the computer-readable storage medium.

[0022] This invention transforms traditional serial data processing into large-scale parallel computing by constructing a three-layer CUDA processing architecture for GPUs. The first layer, a data preprocessing grid, performs parallel cleaning of multi-agent negotiation data; the second layer, a negotiation analysis grid, runs an agent interaction recognition model for parallel pattern recognition; and the third layer, a visualization computing grid, generates visualization data in parallel, significantly improving data processing throughput. This invention employs a lightweight neural network model based on the Transformer architecture, using a multi-head attention mechanism to process negotiation sequences in parallel. The number of attention heads is dynamically adjusted according to the number of agents, fully utilizing the advantages of GPU parallel computing to achieve efficient processing of large-scale multi-agent negotiation data. This invention combines an LSTM-CNN fusion architecture with a scheduling process mapping model, achieving real-time mapping of complex scheduling processes through parallel temporal data processing and spatial relationship calculation, effectively solving the problem of insufficient computing power in traditional processing architectures. In summary, this invention solves the technical problem mentioned in the background art of the inability to achieve real-time parallel processing of high-frequency multi-agent negotiation data. Attached Figure Description

[0023] Figure 1 This is a flowchart of the method of the present invention.

[0024] Figure 2 This is a time-series variation diagram of the multi-agent negotiation intensity in Example 2. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0026] like Figure 1 The diagram shown is a flowchart of a production scheduling process visualization method integrating AI multi-agent and digital twin technologies, provided by the first aspect of this invention. This method includes the following steps:

[0027] S01. Establish a distributed data acquisition layer, and collect workshop equipment status data, material location data, and multi-agent negotiation data at a frequency of 1000Hz through a 5G sensor network. Use a Kalman filter to remove sensor noise from the workshop equipment status data, material location data, and multi-agent negotiation data, and convert the heterogeneous data into a unified spatiotemporal data format and store it in a time-series database.

[0028] S02. Start the three-layer CUDA processing architecture of the GPU. The first layer data preprocessing grid is responsible for real-time data cleaning and feature extraction of the spatiotemporal data format. The second layer negotiation analysis grid runs the agent interaction recognition model to process the multi-agent negotiation sequence in the multi-agent negotiation data. The third layer visualization computing grid runs the scheduling process mapping model to generate visualization data.

[0029] S03. Construct a multi-agent negotiation state graph, and use a force-oriented layout algorithm to map agent nodes to node radii according to the bid price. The bid price comes from the multi-agent negotiation sequence. Use a dynamic clustering algorithm to merge and display similar interactive behaviors, and combine a time axis controller to realize segmented playback of the negotiation process.

[0030] S04. Establish a dual-engine rendering system that synchronizes virtual and real rendering. The real-time rendering engine uses a dynamic LOD algorithm to adjust the display precision according to the viewing distance, which is in the range of (0, 100]m. The offline precision engine generates millimeter-level playback data based on ray tracing technology. The data linkage between the two engines is realized through a spatiotemporal indexing mechanism.

[0031] S05. An abnormal diffusion visualization early warning mechanism is adopted to construct an order resource dependency graph, Dijkstra's algorithm is used to calculate the shortest propagation path of abnormal impact, and a particle system is used to simulate the diffusion process. The particle color changes from yellow to red as the diffusion distance increases.

[0032] S06. Implement an adaptive view recommendation algorithm, collect user operation data to construct feature vectors, and use the view personalized recommendation model to output the optimal view combination scheme. When the sampling frequency ∈ [1, 120) Hz, start the cubic spline interpolation reconstruction algorithm to avoid view switching distortion. When the sampling frequency > 240 Hz, use low-pass filtering downsampling to reduce data redundancy.

[0033] S07. Based on the visualized data, optimize the production scheduling decision, integrate the transparency of multi-agent negotiation, the early warning information of abnormal spread, and the user view feedback into the production scheduling adjustment parameters, generate the final production scheduling plan and output it to the workshop execution system, establish a hierarchical processing system for multi-scale negotiation matrix mapping, and realize the multi-level extraction of the multi-agent negotiation sequence through step-by-step mapping from the equipment-level scale to the workshop-level scale.

[0034] The agent interaction recognition model is structured as a lightweight neural network based on the Transformer architecture, comprising a 4-layer encoder and a 2-layer decoder. It employs a multi-head attention mechanism to process the multi-agent negotiation sequence. The number of attention heads is dynamically adjusted based on the number of currently active agents, calculated as: Number of attention heads = Number of agents / 8 + 2, with a maximum limit of 16 heads. The number of model parameters is controlled within a specified range. Ensure real-time inference performance within the specified timeframe;

[0035] The steps for establishing the training dataset for the agent interaction recognition model include collecting 12 months of historical production scheduling logs, extracting agent ID, interaction timestamp, bidding parameters, and negotiation result fields to construct the multi-agent negotiation sequence, segmenting the data according to a 300-second time window, labeling abnormal and normal negotiation events, and including positive samples in the dataset. negative samples The negative samples were augmented using data augmentation techniques. Maintaining sample balance;

[0036] The training steps of the intelligent agent interaction recognition model include using the AdamW optimizer, setting the learning rate to 0.001, batch size to 32, training for 100 epochs, using the cross-entropy loss function, evaluating the model performance on the validation set every 10 epochs, and triggering an early stopping mechanism when the validation accuracy does not improve for 5 consecutive epochs. Finally, the model achieves a negotiation behavior recognition accuracy of 96.5% on the test set.

[0037] The production scheduling process mapping model is a time-series data processing model based on convolutional neural networks. It integrates LSTM for processing time-series data and CNN for processing spatial relationship data. An attention fusion module is used to combine multi-dimensional features. The spatial attention weight is determined based on three parameters: equipment importance coefficient, material flow frequency, and probability of anomaly occurrence. The weight calculation formula is: Spatial attention weight = ;

[0038] The steps for establishing the training dataset for the production scheduling process mapping model include collecting historical production scheduling data from 50 workshops, including equipment operating parameters, material flow records, and process execution sequences. Training labels are then constructed by combining this data with real-time status data from a digital twin, categorized into three levels based on scheduling complexity: simple, medium, and complex. Each category contains data samples. 1, total One training sample;

[0039] The training steps of the production scheduling process mapping model include a phased training strategy. In the first phase, the LSTM temporal processing module is pre-trained for 50 rounds. In the second phase, the CNN spatial processing module and the LSTM temporal processing module are jointly trained for 100 rounds. In the third phase, the attention fusion module is fine-tuned for 30 rounds. The Adam optimizer is used with an initial learning rate of 0.01, which is decayed to 0.1 times the original rate every 30 rounds. A weighted combination of mean squared error loss function and temporal consistency loss function is used.

[0040] The multi-agent negotiation data includes five categories of data: equipment agent bidding information, order agent demand parameters, material agent inventory status, scheduling agent coordination instructions, and quality agent detection results. Each category of data includes a unique agent identifier, timestamp, status code, parameter value, and priority weight field. During the negotiation process, the agents exchange data through a message passing mechanism.

[0041] The multi-agent system includes an equipment agent responsible for equipment status monitoring and capacity bidding, an order agent responsible for order demand analysis and priority management, a material agent responsible for inventory monitoring and delivery scheduling, a scheduling agent responsible for resource allocation and conflict coordination, and a quality agent responsible for quality detection and anomaly warning. Each agent achieves collaborative optimization of production scheduling decisions through a distributed negotiation algorithm.

[0042] The force-guided layout algorithm is a graph layout algorithm based on physical simulation. It treats agent nodes as point masses and the lines between nodes as springs. It achieves a balance by iteratively calculating the attractive and repulsive forces between nodes, thus realizing an intuitive display of the relationships between agents.

[0043] The dynamic clustering algorithm is an improved K-means clustering method. It dynamically determines the number of cluster centers based on the interaction frequency and behavioral similarity of agents, and groups agents with similar negotiation behaviors into one class to reduce display complexity.

[0044] The dynamic LOD algorithm is an algorithm that dynamically adjusts the display level of detail based on the distance of the observer's viewpoint. The distance ∈ (0, 5]m displays full-precision data, the distance ∈ (5, 20]m displays medium-precision data, and the distance > 20m displays simplified data, thus balancing the display effect and the computing performance.

[0045] Dijkstra's algorithm is a classic algorithm for calculating the single-source shortest path in a graph. It is used to quickly find the shortest propagation path of anomaly effects in the order resource dependency graph, with a time complexity of O(log n). ;

[0046] The personalized view recommendation model is a recommendation system based on a combination of collaborative filtering and deep learning. It predicts the most suitable combination of visualization views by analyzing the user's historical operation behavior, thereby improving the user's operation efficiency.

[0047] The cubic spline interpolation reconstruction algorithm is a numerical method that uses piecewise cubic polynomials to interpolate low-sampling-frequency data, ensuring the continuity and smoothness of the reconstructed signal and avoiding distortion during view switching.

[0048] The specific implementation methods of the above steps are described in detail below.

[0049] The specific implementation of step S01 involves establishing a hierarchical data acquisition system. First, a 5G high-frequency sensor network is deployed at key equipment locations in the workshop. Sensor nodes include vibration sensors, temperature sensors, position sensors, and communication modules. The sampling frequency is set to 1000Hz to ensure high temporal resolution of the data acquisition. During data acquisition, a distributed protocol stack is used to process multi-source heterogeneous data, and a timestamp synchronization mechanism ensures the temporal consistency of data from different sensors. To eliminate sensor noise interference, a Kalman filter algorithm is introduced to preprocess the raw data. This algorithm, based on a state-space model, iteratively estimates the true state value through two stages: prediction and update, effectively filtering out high-frequency noise and random errors. The heterogeneous data standardization process includes data format conversion, unit unification, and spatiotemporal coordinate system alignment. Finally, the processed data is stored in a time-series database, which uses a columnar storage structure to optimize the query performance of the time-series data.

[0050] The specific implementation of step S02 involves constructing a GPU-accelerated parallel computing architecture, which implements a three-layer gridded processing based on the CUDA programming model. The first layer, the data preprocessing grid, is responsible for real-time data cleaning and feature extraction. It employs a parallel stream processing mechanism to deduplicate, detect outliers, and impute missing values ​​in time-series data. The feature extraction module uses a sliding time window method to calculate statistical and frequency domain features. The second layer, the negotiation analysis grid, runs an agent interaction recognition model. This model, based on a transformer architecture, processes multi-agent negotiation sequences and improves model processing speed through multi-threaded parallel inference. The third layer, the visualization computing grid, runs a scheduling process mapping model, converting digital scheduling data into a three-dimensional visualization representation, including geometric modeling, material rendering, and animation generation. This three-layer architecture achieves efficient data transfer through GPU memory management mechanisms, avoiding the CPU-GPU data copy bottleneck.

[0051] The specific implementation of step S03 involves constructing a graph theory-based multi-agent relationship visualization system and employing a force-directed layout algorithm to automatically place agent nodes. This algorithm, based on physical simulation principles, treats agent nodes as charged particles and calculates the optimal node position through the balance of Coulomb and Hooke forces. The node radius is dynamically adjusted according to the bidding price, with three price thresholds: low price corresponds to a small node radius of 2-5 pixels, medium price to a medium radius of 6-10 pixels, and high price to a large radius of 11-15 pixels. A dynamic clustering algorithm clusters agents based on interaction frequency and behavioral similarity, using cosine distance for similarity calculation. The number of cluster centers is adaptively adjusted based on data density, with a density threshold set to 0.3. The timeline controller supports segmented playback of the negotiation process, with a time slice length set to 30 seconds. Users can control the playback speed and positioning via a time slider.

[0052] The specific implementation of step S04 involves establishing a dual-engine collaborative rendering mechanism. The real-time rendering engine employs a dynamic level of detail algorithm to adaptively adjust the model precision based on the observer's viewing distance. Viewing distance calculation is based on the Euclidean distance between the camera position and the target object position. When the viewing distance is within 0 to 5 meters, a full-precision model is displayed, containing complete geometric details and high-resolution textures. When the viewing distance is within 5 to 20 meters, a medium-precision model is displayed, with a 50% reduction in the number of geometric faces and a one-level reduction in texture resolution. When the viewing distance exceeds 20 meters, a simplified model is displayed, using proxy geometry and low-resolution textures. The offline precision calculation engine generates high-precision playback data based on ray tracing algorithms and uses Monte Carlo sampling to calculate global illumination effects with millimeter-level precision. The spatiotemporal indexing mechanism adopts a hybrid structure of spatial hash tables and temporal index trees, supporting rapid retrieval and synchronous updates of data from both engines.

[0053] The specific implementation of step S05 involves establishing an anomaly propagation visualization and early warning system. First, an order-resource dependency graph is constructed, where nodes represent orders and resource entities, and edges represent dependencies and constraints. Dijkstra's shortest path algorithm is used to calculate the propagation path of the anomaly's impact. The algorithm, based on a greedy strategy, gradually expands the path search range, with a time complexity of the square of the number of vertices. Anomaly propagation visualization uses a particle system simulation. The particle emitter is located at the anomaly source, and particles propagate along the dependency path. The propagation speed is adjusted according to the impact intensity, with an intensity threshold set between 0.1 and 1.0. Particle color is mapped using a heatmap; yellow indicates a diffusion distance of 0, and the color gradually transitions to red as the distance increases, with the maximum diffusion distance corresponding to dark red. The particle lifespan is set to 10 to 30 seconds to ensure the duration of anomaly impact visualization.

[0054] The specific implementation of step S06 involves establishing an adaptive view recommendation system. This system records operation sequences, including viewpoint switching, zooming, timeline adjustment, and interactive selection, through a user behavior data acquisition module. User feature vectors are constructed based on three dimensions: operation frequency, dwell time, and interaction depth. Principal component analysis (PCA) is used for feature dimensionality reduction, retaining principal components with a cumulative variance contribution rate exceeding 90%. The personalized view recommendation model is based on a deep neural network architecture, combined with a collaborative filtering mechanism to predict the optimal view combination. The model input consists of user feature vectors and historical preference data, and the output is a view recommendation score. When the data sampling frequency is below 120Hz, a cubic spline interpolation algorithm is used to reconstruct missing frames. Interpolation nodes are selected based on a uniform time interval distribution to ensure the smoothness of the reconstructed signal. When the sampling frequency exceeds 240Hz, a Butterworth low-pass filter is used for downsampling, with a cutoff frequency set at 120Hz and a filter order of 4.

[0055] The specific implementation of step S07 involves integrating multi-dimensional data to optimize production scheduling decisions, using multi-agent negotiation transparency, anomaly propagation warning information, and user view feedback as decision input parameters. Negotiation transparency is quantified based on the integrity of agent interactions and the interpretability of decisions; the transparency threshold is set at 0.7, and negotiations below this value require manual intervention. Anomaly propagation warning information includes three indicators: impact range, propagation speed, and severity; a rescheduling process is triggered when the impact range exceeds 30% of orders. User view feedback is quantified through satisfaction scores and operational efficiency indicators; the view recommendation strategy is adjusted when the satisfaction score is below 3. The multi-scale negotiation matrix is ​​established based on a hierarchical decomposition method, aggregating agent negotiation results layer by layer from equipment level, production line level to workshop level, with different weight coefficients at each level: 0.5 for equipment level, 0.3 for production line level, and 0.2 for workshop level. The final production scheduling plan is generated through a weighted optimization algorithm, with optimization objectives including production efficiency, resource utilization, and cost control.

[0056] Further explanation is needed regarding the agent interaction recognition model, which employs a lightweight neural network structure based on a transformer architecture. This structure includes an input embedding layer, a four-layer encoder, a two-layer decoder, and an output classification layer. The input embedding layer converts the multi-agent negotiated sequence into a high-dimensional vector representation, using a positional encoding mechanism to preserve temporal information. The encoder layer combines a multi-head self-attention mechanism with a feedforward neural network. The multi-head attention mechanism enables parallel computation of the relationships between different positions. The number of attention heads is dynamically adjusted based on the number of currently active agents, calculated as the number of agents divided by 8 plus 2, with a maximum limit of 16 heads to ensure manageable computational complexity. The decoder layer uses a masked self-attention mechanism to prevent information leakage and fuses the encoder output through a cross-attention mechanism. The model's parameter count is strictly controlled to within ten million, achieving model lightweighting through parameter sharing and network pruning techniques.

[0057] The training dataset for the intelligent agent interaction recognition model is built through four stages: historical data collection, data preprocessing, feature engineering, and annotation. The historical data collection stage extracts 12 months of production scheduling logs from the production management system, including key fields such as agent identifiers, interaction timestamps, bidding parameters, and negotiation results. The data preprocessing stage cleans and formats the raw logs, removing invalid records and abnormal data, and standardizing the data format and encoding. The feature engineering stage segments the data according to 300-second time windows, extracting features such as negotiation frequency, bid change rate, and response time. The annotation stage involves domain experts classifying negotiation events, labeling normal negotiation events as 1 and abnormal negotiation events as 0, constructing a dataset containing 150,000 positive samples and 30,000 negative samples. To address the sample imbalance problem, synthetic minority class oversampling is used to expand the negative samples to 60,000, while data augmentation methods such as random rotation and time warp are used to improve the model's generalization ability.

[0058] Further explanation is needed regarding the production scheduling process mapping model, which employs a multimodal deep learning architecture, integrating a Long Short-Term Memory (LSTM) network for processing temporal data and a Convolutional Neural Network (CNN) for processing spatial relationship data. The temporal processing branch uses a bidirectional LSM network structure with 128 hidden units, capable of capturing temporal dependencies in the production process. The spatial processing branch uses a two-dimensional CNN with three convolutional layers and two pooling layers, a 3×3 kernel size, a stride of 1, and a modified linear unit (MRU) activation function. The attention fusion module uses a self-attention mechanism to weightedly fuse spatiotemporal features. Spatial attention weights are calculated based on equipment importance coefficients, material flow frequency, and anomaly probability. The weight calculation is based on a weighted average method, with equipment importance coefficients accounting for 40%, material flow frequency for 30%, and anomaly probability for 30%. The fused features are mapped to the final visualized data output through a fully connected layer.

[0059] The training dataset for the production scheduling process mapping model employs a hierarchical sampling strategy, collecting historical production scheduling data from 50 workshops of different sizes and types. This data includes multi-dimensional information such as equipment operating parameters, material flow records, and process execution sequences. Data standardization processes include missing value imputation, outlier detection, and data normalization to ensure the comparability of data from different workshops. Based on scheduling complexity, the data is categorized into simple, medium, and complex categories. Complexity assessment is based on the number of equipment, process steps, and constraints. Each category contains 20,000 training samples, totaling 60,000 training samples. Training labels are constructed using real-time digital twin status data, including key information such as equipment status, material location, and process progress. A time alignment algorithm is used to ensure the accuracy of the label data.

[0060] The collaborative work of the three-layer CUDA processing architecture, the agent interaction recognition model, and the scheduling process mapping model has brought about significant technical effects and advancements. First, the GPU parallel computing architecture fully leverages the massively parallel computing capabilities of modern graphics processing units, achieving efficient pipelined processing of data preprocessing, agent negotiation analysis, and visualization computation through a three-layer gridded design, significantly improving the overall computational performance of the system. Second, the agent interaction recognition model based on the converter architecture accurately captures complex interaction patterns between multiple agents through a multi-head attention mechanism. The dynamic adjustment of the number of attention heads allows the model to adapt to agent systems of different scales, improving its versatility and robustness. Third, the scheduling process mapping model, by integrating long short-term memory networks and convolutional neural networks, considers both the temporal characteristics and spatial relationships of the production process. The attention fusion mechanism dynamically adjusts weights based on the importance of actual production factors, making the visualization results more consistent with actual production logic and improving the effectiveness of decision support. The synergistic effect of these three components achieves end-to-end optimization from data acquisition to intelligent analysis to visualization, constructing an efficient, accurate, and real-time intelligent manufacturing visualization decision support system.

[0061] It should be noted that the key technical ideas of this invention include real-time visualization of the multi-agent negotiation process, a GPU three-layer parallel processing architecture, a dual-engine collaborative rendering mechanism, and an adaptive view recommendation algorithm. The real-time visualization technology of the multi-agent negotiation process transforms the abstract agent interaction process into an intuitive visual representation by constructing a negotiation state graph and a force-directed layout algorithm. Compared to traditional static report displays, this technology can reflect the dynamic changes of the negotiation process in real time, helping decision-makers quickly understand complex multi-party negotiation relationships and significantly improving decision-making efficiency and accuracy. The GPU three-layer parallel processing architecture addresses the performance bottleneck of traditional CPU serial processing by parallelizing data preprocessing, negotiation analysis, and visualization computation through the CUDA programming model, fully leveraging the advantages of GPUs in large-scale parallel computing. Compared to traditional single-threaded processing, this architecture can process multiple data streams simultaneously, significantly reducing data processing latency. The dual-engine collaborative rendering mechanism combines the advantages of real-time rendering and offline precision computing, achieving a balance between display quality and computational performance through dynamic level of detail algorithms and spatiotemporal indexing mechanisms. Compared to traditional single-rendering methods, this mechanism ensures both smooth real-time interaction and provides high-precision data playback capabilities. The adaptive view recommendation algorithm, based on user behavior analysis and deep learning technology, can automatically recommend the optimal view combination according to the user's operating habits and preferences. Compared with the traditional fixed view mode, this algorithm significantly improves user experience and operational efficiency. The synergistic effect of these four key technological ideas constructs a highly intelligent, user-friendly, and powerful visualization system, achieving a comprehensive technological upgrade from passive display to proactive recommendation, from static analysis to dynamic interaction, and from single processing to parallel computing. This provides strong technical support for production decision-making in intelligent manufacturing environments.

[0062] It should be noted that this invention also solves the following technical problem: the technical problem of multi-scale heterogeneous data fusion processing. In existing multi-agent scheduling systems, data generated by different types of agents have different time scales, data formats, and semantic features. Traditional data processing methods are difficult to effectively integrate these heterogeneous data sources, resulting in a lack of a global optimization perspective in negotiation decisions. This invention establishes a hierarchical processing system for multi-scale negotiation matrix mapping, achieving step-by-step mapping from the equipment-level scale to the workshop-level scale. Combined with the spatial attention mechanism of the scheduling process mapping model, weight allocation is determined based on three parameters: equipment importance coefficient, material flow frequency, and anomaly probability, effectively solving the problem of unified modeling and fusion processing of multi-scale heterogeneous data. Simultaneously, this invention also solves the technical problem of balancing real-time processing of high-frequency data streams with visualization rendering performance. Traditional visualization systems often experience rendering delays and display stuttering when processing 1000Hz high-frequency sampled data. This invention uses a dynamic LOD algorithm to adjust display precision based on viewing distance, and employs a cubic spline interpolation reconstruction algorithm and low-pass filtering downsampling processing technology to achieve an optimized balance between visualization quality and real-time rendering performance.

[0063] A second aspect of the present invention provides a computer-readable storage medium storing program instructions, which, when executed in a computer, are used to perform the above-described method for visualizing the scheduling process by integrating AI multi-agent and digital twin technologies.

[0064] A third aspect of the present invention provides a production scheduling process visualization system integrating AI multi-agent and digital twin, comprising the aforementioned computer-readable storage medium. The system is any one of a computer, a server, or a microcontroller. The computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes the program instructions stored in the computer-readable storage medium.

[0065] Specifically, the principle of this invention is as follows: The fundamental principle that enables this invention to solve the problem of real-time parallel processing of high-frequency negotiation data from multiple agents lies in the construction of a three-layer CUDA parallel computing architecture on a GPU, which fully utilizes the massive parallel processing capabilities of the GPU to achieve efficient data processing. At the architecture design level, the first-layer data preprocessing grid transforms traditional serial data cleaning into thousands of parallel threads executing simultaneously. Each thread is responsible for processing independent data blocks. Data access patterns are optimized through shared memory and texture memory, significantly improving data cleaning efficiency. The second-layer negotiation analysis grid runs an agent interaction recognition model. Based on the parallel computing characteristics of the Transformer architecture, the multi-head attention mechanism achieves fully parallel computation on the GPU. The strategy of dynamically adjusting the number of attention heads according to the number of agents ensures reasonable allocation of computing resources, and the number of model parameters is controlled within 10 to the power of 7 to guarantee real-time inference performance. The third-layer visualization computing grid runs a scheduling process mapping model. LSTM temporal processing and CNN spatial processing are executed in parallel on the GPU. The attention fusion module quickly calculates spatial attention weights through parallel matrix operations, achieving efficient mapping of complex spatiotemporal data. At the computational optimization level, the CUDA processing architecture employs a strategy combining pipelined parallelism and data parallelism. It hides memory access latency through asynchronous data transfer and computational overlap techniques, leveraging the high bandwidth of the GPU to achieve rapid exchange of large-scale data. Through reasonable task allocation and load balancing strategies, the entire architecture transforms multi-agent negotiation data, which originally required serial processing, into large-scale parallel computing tasks, enabling real-time processing of 1000Hz high-frequency negotiation data and effectively overcoming the technical bottleneck of insufficient computing power in traditional CPU architectures.

[0066] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0067] The specific implementation of step S01 involves establishing a hierarchical data acquisition system, collecting workshop data through a 5G high-frequency sensor network, and performing data preprocessing using a Kalman filter. The state prediction equation of the Kalman filter algorithm is expressed as follows:

[0068] ;

[0069] ;

[0070] In the formula, This is the predicted state value at time k, in sensor measurement units. The state transition matrix is ​​dimensionless. This is the optimal estimate at time k-1; To control the input matrix; For control vectors; The prediction error covariance matrix; This is the transpose of the state transition matrix; This is the process noise covariance matrix, with a default value of [value to be filled in]. The Kalman filter update equation is expressed as follows:

[0071] ;

[0072] ;

[0073] In the formula, The Kalman gain matrix is ​​dimensionless. The observation matrix; This is the transpose of the observation matrix; To observe the noise covariance matrix, the empirical value is... ; Let be the observation value at time k; This is the optimal estimate at time k.

[0074] The specific implementation of step S02 is the same as described above, and will not be repeated in detail here.

[0075] The specific implementation of step S03 involves constructing a multi-agent negotiation state graph and using a force-directed placement algorithm to achieve automatic placement of agent nodes. The node position update equation of the force-directed placement algorithm is expressed as follows:

[0076] ;

[0077] ;

[0078] In the formula, Let be the resultant force vector between node i and node j, in Newtons; This is the Coulomb repulsion coefficient, which is 100 by default. This is the Hooke attraction coefficient, with a default value of 0.1. , Let i be the position vectors of nodes i and j; The ideal distance between nodes; Let be the velocity vector of node i at time t; Let be the velocity vector of node i at time t+1; This represents the node quality, typically set to 1. This is the time step, which defaults to 0.01 seconds. The similarity calculation formula for the dynamic clustering algorithm is as follows:

[0079] ;

[0080] In the formula, Let be the cosine similarity between agents i and j, which is dimensionless and takes values ​​in the range [-1, 1]. Let be the standardized value of agent i on the p-th dimension feature; The number of feature dimensions.

[0081] The specific implementation of step S04 involves establishing a dual-engine rendering system, where the dynamic level of detail algorithm adjusts the display precision based on the observer's viewing distance. The viewing distance calculation formula is as follows:

[0082] ;

[0083] ;

[0084] In the formula, The Euclidean distance from the observer to the target object, in meters; The coordinates of the camera position; The coordinates of the target object's position; The level of detail coefficient is dimensionless. The maximum line-of-sight distance is set to 100 meters by default.

[0085] The specific implementation of step S05 involves constructing an anomaly propagation visualization and early warning mechanism, and using the Dijkstra algorithm to calculate the propagation path of anomaly impacts. The Dijkstra shortest path calculation formula is expressed as follows:

[0086] ;

[0087] ;

[0088] In the formula, The weight of the shortest distance from the source node to node v is dimensionless. The weight is the shortest distance from the source node to node u; Let be the edge weight from node u to node v; The set of nodes for which the shortest distance has been determined; The color value of the particle at time t; The starting color value corresponds to the RGB value of yellow (255, 255, 0); The terminating color value corresponds to the red RGB value (255, 0, 0); This represents the current diffusion distance. The maximum diffusion distance is usually taken as the diameter of the graph.

[0089] The specific implementation of step S06 is to implement the adaptive view recommendation algorithm. The formula for dynamically adjusting the number of attention heads in the agent interaction recognition model is expressed as follows:

[0090] ;

[0091] In the formula, The number of attention heads is a dimensionless integer. This represents the number of currently active agents. This is the floor function. The formula for constructing user feature vectors is as follows:

[0092] ;

[0093] In the formula, The user feature vector is dimensionless. This refers to the operation frequency, measured in times per minute. Maximum operating frequency; The dwell time is in seconds. This is the maximum stay time; The interaction depth is dimensionless. The maximum interaction depth. The formula for calculating the view personalization recommendation score is as follows:

[0094] ;

[0095] In the formula, A rating is recommended for the view; it is dimensionless and ranges from [0, 1]. Let k be the similarity weight for user k. Cosine similarity between user feature vectors; For user k, the historical ratings of a specific view; For reference, the number of users is typically set to 50. The interpolation function of the cubic spline interpolation reconstruction algorithm is expressed as follows:

[0096] ;

[0097] In the formula, Let j be the cubic spline function of the j-th segment; Let x be the x-coordinate of the j-th interpolation node; , , , These are undetermined coefficients, determined through boundary conditions; The x-coordinate of the interpolation point, with a range of values. The transfer function of the Butterworth low-pass filter when the sampling frequency exceeds 240Hz is expressed as follows:

[0098] ;

[0099] In the formula, The filter transfer function is dimensionless. This is the gain constant, which defaults to 1. It is a complex frequency variable; Let r be the r-th pole; This is the filter order, which defaults to 4.

[0100] The specific implementation of step S07 is to optimize production scheduling decisions based on visualized data. The formula for calculating spatial attention weight is as follows:

[0101] ;

[0102] In the formula, represents the spatial attention weight, which is dimensionless and takes values ​​in the range [0, 1]. This represents the equipment importance coefficient. This represents the maximum equipment importance coefficient. This refers to the material turnover frequency, measured in times per hour. This represents the maximum material turnover frequency. This represents the probability of an anomaly occurring. This represents the maximum probability of an anomaly occurring. , , Assign coefficients to the weights. The objective function of the weighted optimization algorithm is expressed as follows:

[0103] ;

[0104] In the formula, To optimize the objective function value, it is dimensionless; Production efficiency is expressed in units per hour; To maximize production efficiency; Resource utilization rate, dimensionless; To maximize resource utilization; Production cost, in yuan; To maximize production cost; , , To optimize the weighting coefficients, empirical values ​​of 0.5, 0.3, and 0.2 are used. The hierarchical mapping formula for the multi-scale negotiation matrix is ​​expressed as follows:

[0105] ;

[0106] In the formula, A workshop-level negotiation matrix; , , These are negotiation matrices at the equipment level, production line level, and workshop level, respectively. , , These represent the maximum values ​​of each negotiation matrix level; , , This represents the hierarchical weight coefficient.

[0107] The Kalman filter algorithm, through recursive calculations in two stages—state prediction and observation update—can provide optimal estimation of the system state in the presence of process noise and observation noise. Compared with the traditional moving average filtering method, this algorithm can effectively handle non-stationary noise environments, improve the accuracy and stability of sensor data, and provide a high-quality data foundation for subsequent agent negotiation analysis.

[0108] Force-directed layout algorithms are based on physical simulation principles. They achieve equilibrium by simulating the Coulomb repulsion and Hooke attraction between nodes, where the Coulomb repulsion term... To ensure proper spacing between nodes, the Hooke attraction term... Maintain the connections between nodes.

[0109] ;

[0110] In the formula, is the Euclidean distance between nodes i and j, in pixels; and These are the two-dimensional coordinates of nodes i and j, respectively. Compared to traditional hierarchical layout methods, this algorithm can adaptively adjust node positions, enabling an intuitive display of agent relationships and improving the visualization and user understanding of the negotiation process.

[0111] Dijkstra's shortest path algorithm gradually expands the search range through a greedy strategy to ensure that the shortest path from the source node to all other nodes is found. Compared with methods such as breadth-first search, this algorithm has higher computational efficiency in weighted graphs, providing an algorithmic guarantee for the rapid identification of the propagation path of anomalies.

[0112] The dynamic adjustment formula for the number of attention heads adaptively determines model parameters based on the current number of active agents. The linear mapping relationship enables dynamic allocation of the number of heads.

[0113] ;

[0114] In the formula, It is a set of integers; For integer variables. Compared to methods with a fixed number of attention heads, this formula can balance model complexity and computational efficiency, controlling computational resource consumption while ensuring the accuracy of negotiation behavior recognition, thus improving the real-time performance of the system.

[0115] Personalized view recommendation ratings are calculated using a collaborative filtering mechanism. To achieve accurate prediction of user preferences, cosine similarity calculation is used. This ensures an accurate measurement of the similarity between user feature vectors.

[0116] ;

[0117] ;

[0118] In the formula, The l-th component of the user feature vector; Let l be the l-th component of the feature vector of user k. Compared with content-based recommendation methods, this recommendation mechanism can better capture users' implicit preferences, thereby improving the accuracy of view recommendations and user satisfaction.

[0119] The spatial attention weight calculation formula comprehensively considers three key factors: equipment importance, material flow frequency, and probability of anomaly occurrence. It determines the attention allocation through a weighted average method. Compared with the uniform weight allocation method, this formula can dynamically adjust the attention focus according to the importance of actual production factors, making the visualization results more consistent with the actual production logic and improving the pertinence and effectiveness of decision support.

[0120] The objective function of the weighted optimization algorithm is obtained through The multi-objective optimization structure achieves synergistic optimization of production efficiency, resource utilization, and cost control, with the cost item employing... The reverse normalization form ensures that the lower the cost, the larger the objective function value. Compared with single-objective optimization methods, this function can find the optimal solution under multiple constraints, improving the overall performance and practicality of the production scheduling scheme.

[0121] To better understand and implement this invention, the following is a specific application scenario of this invention, Example 2:

[0122] A technical team implemented the AI ​​multi-agent and digital twin-integrated production scheduling visualization method of this invention in an automotive parts manufacturing workshop. The workshop mainly produces core components such as engine blocks, transmission housings, and brake discs, and includes 45 CNC machining centers, 23 assembly stations, and 12 automated production lines.

[0123] In step S01, the team deployed 268 5G sensor nodes in the workshop, using a 1000Hz high-frequency acquisition mode to monitor the equipment's operating status in real time. The sensor network covered all key equipment, collecting parameters such as spindle speed, tool wear, temperature changes, and vibration amplitude. Material position sensors used RFID and visual recognition technology to track the position information and processing progress of 3580 workpieces in real time. The multi-agent negotiation data acquisition frequency was 50Hz, recording bidding information, negotiation instructions, and decision results among the agents. A Kalman filter was used to denoise the raw data, with filter parameters Q set to 0.001 and R set to 0.1, effectively removing high-frequency noise and measurement errors. The processed data was uniformly converted into a standard format containing timestamps, location coordinates, device IDs, and status values, and stored in a 50TB time-series database.

[0124] Step S02 initiates a CUDA processing architecture equipped with four Tesla V100 GPUs. The first-layer data preprocessing grid contains 512 processing units, performing real-time cleaning of spatiotemporal data to remove outliers and duplicates, while extracting key information such as equipment operation trends, material flow patterns, and abnormal event characteristics. The second-layer negotiation analysis grid runs an intelligent agent interaction recognition model, which includes a 4-layer encoder and a 2-layer decoder. The encoder dimension is 256, and the decoder dimension is 128. Based on the current number of 32 active agents, the number of attention heads is calculated to be 6, and the model parameter count is controlled within a certain range. Each layer ensures a single inference time of less than 15ms. The third layer is a visualization computing grid running a production scheduling process mapping model. This model integrates LSTM and CNN structures, with 128 LSTM units and a 3×3 CNN convolutional kernel size, generating multi-dimensional visualization data that includes equipment status, material location, and negotiation process.

[0125] In step S03, the constructed multi-agent negotiation state graph contains 32 agent nodes. A force-directed layout algorithm is used to map the bidding price range of the equipment agents (580-1250 yuan) to node radii of 2-8 pixels, and the urgency of the order agents is mapped to the color depth of the nodes. A dynamic clustering algorithm merges agents with similar interactive behaviors, identifying five main negotiation groups: high-precision processing group, batch production group, urgent order group, quality inspection group, and material delivery group. The timeline controller supports 1x, 2x, and 4x playback speeds, enabling segmented playback of the negotiation process and location of key events.

[0126] In the virtual-real synchronous dual-engine rendering system established in step S04, the real-time rendering engine dynamically adjusts the display precision based on the user's viewpoint distance. When the viewing distance is within 0-5m, a full-precision 3D model of the device is displayed, including all detailed features and texture maps. When the distance is 5-20m, a medium-precision model is displayed, reducing the number of polygons to 60% of the original. When the distance exceeds 20m, a simplified model and procedural textures are used, reducing the number of polygons to 20% of the original. The offline precision engine generates playback data with millimeter-level precision based on ray tracing technology, with a single frame rendering time of 850ms, achieving smooth playback through pre-calculation and caching mechanisms. The spatiotemporal indexing mechanism adopts an octree structure, supporting fast spatial queries and temporal positioning.

[0127] Step S05's anomaly propagation visualization and early warning mechanism constructs a dependency graph containing 178 order nodes and 245 resource dependencies. Dijkstra's algorithm is used to calculate the anomaly impact path; when a device failure is detected, the algorithm calculates the shortest propagation path within 3.2ms. A particle system simulates the diffusion process, with an initial particle count of 500, a particle lifespan of 8 seconds, and a color that gradually changes from initial yellow (RGB(255, 255, 0)) to red (RGB(255, 0, 0)). The diffusion speed is set to 0.5-2.0 m / s based on the tightness of the dependencies, providing an intuitive display of the impact range.

[0128] In step S06, the adaptive view recommendation algorithm collected historical operation data from 15 operators and constructed a vector containing 12 features, including view switching frequency, dwell time, zoom level, and rotation angle. The personalized view recommendation model uses a collaborative filtering algorithm, and the training dataset includes... The operation record shows that the model accuracy reached 89.5%. When the sensor sampling frequency is 60Hz, a cubic spline interpolation reconstruction algorithm is activated with an interpolation node interval of 16.7ms to ensure smooth view switching. When the sampling frequency exceeds 240Hz, a low-pass filter with a cutoff frequency of 120Hz is used for downsampling to reduce data redundancy by 85%.

[0129] Step S07 optimizes production scheduling decisions based on visualized data. The transparency of multi-agent negotiation is set to 0.75, the weight of anomaly propagation warning information is 0.4, and the weight of user view feedback is 0.3, integrated into production scheduling adjustment parameters. The established multi-scale negotiation matrix mapping system includes four levels: equipment level, process level, production line level, and workshop level, with negotiation cycles of 5 minutes, 15 minutes, 1 hour, and 4 hours for each level, respectively. The final generated production schedule includes detailed scheduling arrangements for 168 processing tasks, with an estimated completion time of 72 hours and a resource utilization rate of 87.3%.

[0130] The intelligent agent interaction recognition model was trained using historical data from the workshop over the past 12 months. The dataset contains positive samples. The list covers scenarios such as normal negotiation, successful bidding, and resource allocation. Negative samples have been expanded using data augmentation techniques. The model identifies negotiation behavior scenarios including failure, resource conflicts, and abnormal interruptions. The model was trained using the AdamW optimizer with an initial learning rate of 0.001 and a batch size of 32. After 100 training rounds, the model achieved a negotiation behavior recognition accuracy of 96.5% and a recall of 94.8% on the test set.

[0131] The training data for the production scheduling process mapping model comes from historical production scheduling data of this workshop and 49 other similar workshops. The total training set... The samples are divided into three categories according to the production scheduling complexity, as shown in Table 1:

[0132] Table 1. Statistics on Production Scheduling Complexity

[0133]

[0134] The model training employed a phased strategy: the first phase involved 50 epochs of pre-training the LSTM module, the second phase involved 100 epochs of joint training, and the third phase involved 30 epochs of fine-tuning. The final model achieved a scheduling accuracy of 92.3% on the test set, with a mean absolute error of 8.5%.

[0135] The specific composition of the multi-agent negotiated data is shown in Table 2:

[0136] Table 2. Data Structure for Multi-Agent Negotiation

[0137]

[0138] Each intelligent agent exchanges data through a message queue-based communication mechanism, with message transmission latency controlled within 5ms. The equipment intelligent agent monitors the operating status of 45 pieces of equipment and updates capacity bidding information in real time. The order intelligent agent manages 168 work-in-process orders, dynamically adjusting priority weights. The material intelligent agent tracks the inventory status and delivery scheduling of 3580 workpieces. The scheduling intelligent agent is responsible for resource conflict coordination, processing an average of 23 coordination requests per hour. The quality intelligent agent integrates data from 18 inspection stations, enabling anomaly warnings and quality traceability.

[0139] The system's overall architecture adopts a layered design, with data acquisition, processing and analysis, visualization and rendering, and decision application layers each performing their respective functions, and data flow and command issuance achieved through a high-speed network. For example... Figure 2 As shown, the dynamic evolution of the multi-agent negotiation process exhibits obvious periodicity and suddenness. The negotiation intensity can reach 200 times / hour at its peak and drop to 50 times / hour at its trough.

[0140] The technological advancements of this invention compared to traditional scheduling methods are mainly reflected in the following aspects: Traditional scheduling methods rely on manual experience and static planning, making it difficult to cope with dynamically changing production environments. This invention, however, achieves distributed decision-making through a multi-agent negotiation mechanism, where each agent makes autonomous decisions based on real-time data, reaching a globally optimal solution through negotiation. Traditional methods have limited visualization capabilities, making it difficult for operators to intuitively understand complex scheduling logic. This invention constructs a synchronized virtual and real 3D visualization environment using digital twin technology, transforming the abstract negotiation process into an intuitive graphical interface. Traditional anomaly handling is often a passive response, with the impact spreading by the time a problem is discovered. This invention, through an anomaly propagation visualization and early warning mechanism, can predict the path of anomaly impacts, enabling proactive prevention and control. Traditional systems lack personalized adaptation capabilities. This invention, through an adaptive view recommendation algorithm, provides customized visualization solutions based on different users' operating habits, improving the efficiency and user-friendliness of human-computer interaction. These technological advancements fundamentally change the management model of the scheduling process, shifting from static planning to dynamic optimization, from passive response to proactive prevention, and from standardized interfaces to personalized experiences.

[0141] It should be noted that the variables involved in this invention are explained in detail in Tables 3 and 4 below.

[0142] Table 3. Variable Explanation Table (Part 1)

[0143]

[0144] Table 4. Variable Explanation Table (Part Two)

[0145]

[0146] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for visualizing the production scheduling process by integrating AI multi-agent and digital twin technologies, characterized in that, include: A distributed data acquisition layer is established to collect workshop equipment status data, material location data, and multi-agent negotiation data through a 5G sensor network. A Kalman filter is used to remove sensor noise from the workshop equipment status data, material location data, and multi-agent negotiation data, and the heterogeneous data is converted into a unified spatiotemporal data format and stored in a time-series database. A three-layer CUDA processing architecture for GPUs is implemented. The first layer, the data preprocessing grid, is responsible for real-time data cleaning and feature extraction of spatiotemporal data formats. The second layer, the negotiation analysis grid, runs an agent interaction recognition model to process multi-agent negotiation sequences in the multi-agent negotiation data. The third layer, the visualization computing grid, runs a production scheduling process mapping model to generate visualization data. A multi-agent negotiation state graph is constructed, and a force-directed layout algorithm is used to map agent nodes to node radii based on bid prices. A dynamic clustering algorithm is used to group similar interactive behaviors for display, and a timeline controller is used to achieve segmented playback of the negotiation process. A virtual-real synchronized dual-engine rendering system is established. The real-time rendering engine uses a dynamic LOD algorithm to adjust the display precision according to the viewing distance, while the offline precision calculation engine generates playback data based on ray tracing technology. A spatiotemporal indexing mechanism is used to achieve data linkage between the two engines. An anomaly diffusion visualization early warning mechanism is adopted, an order resource dependency graph is constructed, the Dijkstra algorithm is used to calculate the shortest propagation path of anomaly impacts, and a particle system is used to simulate the diffusion process. An adaptive view recommendation algorithm is implemented, collecting user operation data to construct feature vectors, and using a view personalized recommendation model to output the optimal view combination scheme. Based on visualized data, production scheduling decisions are optimized by integrating multi-agent negotiation transparency, anomaly propagation early warning information, and user view feedback into production scheduling adjustment parameters, generating the final production scheduling plan and outputting it to the workshop execution system.

2. The method for visualizing the production scheduling process by integrating AI multi-agent and digital twin technologies according to claim 1, characterized in that, The 5G sensor network collects data at a frequency of 1000Hz and establishes a hierarchical processing system for multi-scale negotiation matrix mapping. The hierarchical mapping from the equipment-level scale to the workshop-level scale enables multi-level extraction of multi-agent negotiation sequences.

3. The method for visualizing the production scheduling process by integrating AI multi-agent and digital twin technologies according to claim 2, characterized in that, The structure of the intelligent agent interaction recognition model is a lightweight neural network based on the Transformer architecture, which includes a 4-layer encoder and a 2-layer decoder. It uses a multi-head attention mechanism to process multi-agent negotiation sequences. The number of attention heads is dynamically adjusted according to the number of currently active intelligent agents. The calculation formula is: number of attention heads = number of intelligent agents / 8 + 2, with a maximum limit of 16 heads.

4. The method for visualizing the production scheduling process by integrating AI multi-agent and digital twin technologies according to claim 3, characterized in that, The production scheduling process mapping model is a time-series data processing model based on convolutional neural networks. It integrates LSTM for processing time-series data and CNN for processing spatial relationship data. It uses an attention fusion module to combine multi-dimensional features. The spatial attention weight is determined based on three parameters: equipment importance coefficient, material flow frequency, and probability of anomaly occurrence.

5. The method for visualizing the production scheduling process by integrating AI multi-agent and digital twin technologies according to claim 4, characterized in that, The steps for establishing the training dataset for the production scheduling process mapping model include collecting historical production scheduling data from 50 workshops, including equipment operating parameters, material flow records, and process execution sequence, and constructing training labels by combining them with real-time status data from digital twins.

6. The method for visualizing the production scheduling process by integrating AI multi-agent and digital twin technologies according to claim 5, characterized in that, The production scheduling process mapping model training steps include adopting a phased training strategy. The first phase pre-trains the LSTM temporal processing module for 50 rounds. The second phase jointly trains the CNN spatial processing module and the LSTM temporal processing module for 100 rounds. The third phase fine-tunes the attention fusion module for 30 rounds, using the Adam optimizer with an initial learning rate of 0.01, which decays to 0.1 times the original rate every 30 rounds.

7. The method for visualizing the production scheduling process by integrating AI multi-agent and digital twin technologies according to claim 6, characterized in that, The multi-agent negotiation data includes five categories of data: equipment agent bidding information, order agent demand parameters, material agent inventory status, scheduling agent coordination instructions, and quality agent detection results. Each category of data includes a unique agent identifier, timestamp, status code, parameter value, and priority weight field.

8. The method for visualizing the production scheduling process by integrating AI multi-agent and digital twin technologies according to claim 7, characterized in that, The multi-agent system includes an equipment agent responsible for equipment status monitoring and capacity bidding, an order agent responsible for order demand analysis and priority management, a material agent responsible for inventory monitoring and delivery scheduling, a scheduling agent responsible for resource allocation and conflict coordination, and a quality agent responsible for quality detection and anomaly warning.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program instructions, which, when executed in a computer, are used to perform the production scheduling process visualization method integrating AI multi-agent and digital twin as described in any one of claims 1-8.

10. A production scheduling process visualization system integrating AI multi-agent and digital twin technologies, characterized in that, The system includes the computer-readable storage medium of claim 9, wherein the system is any one of a computer, a server, or a microcontroller, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes the program instructions stored in the computer-readable storage medium.

Images